Why Aria Is to AI What Arista Was to Cloud | Gavin Baker & Mansour Karam

I watched The Last Samurai recently. I don't know if anybody's seen it, but it's about a master samurai colliding with a machine gun. And I feel like I'm a samurai and I'm not going to lose a fight with a katana. But the machine gun is here and it's AI. And Stefan just said that we need to figure out how to be 1000X investors. So let's add that to the mission statement. No, but seriously, Mansour, I was really excited before. Holy [__] that was awesome. Yeah.

You know, it's actually very interesting. MFU is a training-centric kind of metric. And it goes back to, you know, training does not generate revenue. And I would actually just humb submit, I was just thinking watching this, that hey, we can reduce your cost for training. But more importantly, if we double the utilization of your AI factory, we're literally going to double your revenues. That is the sale on the inference side. Indeed. Like we'll save you money on your training runs, they'll be more reliable, you'll have to checkpoint less, they'll cost less, you'll get more out of your XPUs. But on the inference side, forget everything else, we can materially improve your revenue.

Yeah, I actually think it's when a technological discontinuity encounters first principles thinking and it's just that simple. Well said. Yeah, and so GPUs were a technological discontinuity for graphics for a long time. And then we realized graphics math was very similar to deep learning and kind of Jensen went all in on that. You know, in some ways EVs and Tesla, that was like a technological discontinuity and a lot of the people who were trying to build EVs before were building them the same way and Elon said, 'Wow, that's crazy. Let's put the battery at the bottom of the car. Then we'll have a lower center of gravity. The car will handle better.' First principles thinking. And here we have a technological discontinuity that I don't think anyone has really applied first principles thinking to yet. And I think part of this is because everyone involved in standing up these clusters has been almost too busy to optimize.

Stranded. Stranded. Yeah, I mean there's, you know, if Anthropic's new cloud and cloud providers were using our AM, maybe they wouldn't have 30 billion in revenue, they'd be at 60. And it is literally that order of magnitude of importance. But it also does really feel like the Arista IPO to me. You know, the Arista IPO pitch was really simple. I guess we have some people my age in the room, but everybody thinks iOS is Apple, but the original iOS was Cisco and they just came in and they said, you know, hey, this was built for a different era, the cloud is here, we have EOS, it's purpose-built from first principles, ground up for the age of the cloud, not the age of the enterprise or the early days of what used to be called web 1.0. Right. And this is we're doing the same thing. You know, there's more traffic on the backplane of an NVL-72 than there is over the entire internet. And obviously an operating system built for the former is not going to work for the latter, particularly when they collide, which they do with agents. So I think it's really exciting. I think it's awesome. Yeah, we're actually building the networking company for the AI era.